Bostrom Will Not Pick a Fork: The Simulation Argument Is a Disjunction, Not a Bet

Lex Fridman opens by asking Nick Bostrom, a philosopher at the University of Oxford and director of the Future of Humanity Institute, to play the hit: what is the simulation hypothesis? Bostrom wants the claim taken literally, not as a metaphor that the universe is somehow informational. An advanced civilization has built a lot of computers. What we see, think, and feel — including our own brains — exists because one of those computers is running certain programs. Today's machines would do, he says, if they were bigger. Quantum computers might help, but they are not required for the conclusion that a technologically mature civilization could host conscious beings inside simulations.

He does not pitch the idea as a tool for computer science. It is informative, interesting, maybe important, and of interest to philosophy, cosmology, and physics, because the physics of the computer running the simulation need not match the physics inside it. Then he draws the distinction that most popular versions blur. The *simulation hypothesis* is the claim that we are in a computer simulation. The *simulation argument* is a disjunction: at least one of three propositions is true. The argument, on its own, does not say which.

The first proposition is that almost all civilizations at our current stage of technological development go extinct before they reach technological maturity — a great filter across a vast cosmos. The second is that civilizations that do mature lose interest in creating ancestor simulations. The third is that we are living in a simulation. Fridman notes that people treat the argument and the hypothesis as the same term. Bostrom's whole career of answering this question is, in effect, a refusal to collapse them. He will not pick a favorite fork.

Whether fork one is plausible depends on a concept Bostrom finds useful beyond this argument: technological maturity. Even with current knowledge you can put a lower bound on what a civilization could do by developing technologies we already see are possible. His colleague Eric Drexler, back in the 1980s, used theory and computer modeling to study molecular manufacturing — atomically precise structures nobody could fabricate then or now. You can still say: if you put these atoms together this way, the system is stable, it rotates at this speed, and it has these computational characteristics. Drexler also sketched paths that would, in the fullness of time, get there.

Space colonization supplies another bound. The speed of light is an upper limit; current rockets are a lower one, and you can go faster just by making them bigger and adding fuel. Combine molecular nanotechnology with that spatial reach and you get a very high lower bound on compute. Drexler characterized a crude computer, built from those parts, that would perform at a million times the human brain while being about the size of a sugar cube. He did not claim it was optimal. It is a floor: if you can work at atomic precision, you can at least do that. Maturity, for Bostrom, means you have maxed out — or come within 99.9 percent of — the general-purpose, widely useful technologies that can exist in this universe.

He would guess there is a ceiling, not an endless spring of new basic capabilities, and that a human-originating civilization would reach it fairly shortly after machine superintelligence, on historical timescales rather than millions of years. That guess is independent of the simulation argument. For the argument itself, it does not matter whether the ceiling is 100 years or 50 million, or whether technology goes up forever, so long as we know a certain level is attainable. Hardness *does* matter if you want to *rank* the forks. If simulations are billions of years away, extinction along the way becomes more likely, and probability shifts toward the first alternative. The disjunction still stands either way.

Fork one is not "something could kill us." An asteroid could kill *us*. For the first disjunct to be true, almost all civilizations throughout the cosmos must also fail to reach technological maturity. That quietly assumes there are many of them. On the most popular current cosmological theory, Bostrom says, the universe is infinitely big. Infinite matter with local randomness yields all possible outcomes with probability one, infinitely repeated. Extraterrestrials would then be a certainty, and infinitely many. If we were literally the only civilization, fork one would become trivially true the day we died. If there are many, they would virtually all have to succumb the same way.

That is why he separates two questions people mash together. What is most likely to kill *us* is one inquiry. If something kills *almost everyone, everywhere*, what would that be is another. The second answer has to be a risk factor that is roughly uniform over all possible civilizations — a structural obstacle, not a quirk of human physiology or one technology tree. Fridman jokes that interesting people eventually arrive at the Fermi paradox. Bostrom's angle is less "where is everybody?" and more "what kind of filter could be that fair?"

If we were literally the only civilization, fork one would become trivially true the day we died: a very high fraction of civilizations at our stage would have gone extinct before maturity, because the fraction would be us. The interesting case is the large-numbers case, which popular cosmology now makes the default. That is also why a local disaster is not enough. You would have to postulate something that almost every other intelligent species succumbs to as well, wherever it sits and however it is built. Bostrom's other writing on existential risk is aimed at us. The first fork of the simulation argument is aimed at everyone.

The second proposition is a convergence hypothesis: many civilizations make it to maturity, and *all* of them — not a few, but out of a million maybe not a single one — choose not to run these simulations. "Lose interest" sounds like boredom. Fridman lists harder versions: it could be exceptionally difficult, ethically forbidden, or a change to the fabric of reality. Bostrom rules out mere difficulty. Maturity, in the argument, already means you could create many simulations with a tiny fraction of resources. It would not be a debate about spending half of GDP forever on one run. If a small fraction of a civilization felt like doing it at any point during millions of years, there would be millions of simulations. So the second fork needs a very strong, shared reason not to.

An ancestor simulation, as he uses the term, is not a video-game NPC with an avatar and a tiny script. It contains people like those we think lived on Earth, with brains simulated at enough granularity that they have the same subjective experiences we do. Later in the conversation he returns to why this fork is the foggiest. It does not seem implausible to him that the leap to maturity — almost certainly via machine superintelligence, before anyone can mint large numbers of ancestor simulations — could radically change what a civilization is optimizing for.

Fridman still hears "bored." Bostrom's bar is higher than a mood. Ethical refusals could do the work, but difficulty cannot, once you are mature: the simulations would be cheap. The second fork is therefore a claim about values and incentives after a civilization can already do the thing, not about engineering failure on the way there. That is why he later treats this alternative as the one we understand least. From our present vantage, running ancestor simulations looks like a free choice some fraction of people would take. From a mature vantage, it might look obligatory — or obviously off-limits.

The argument smuggles in a premise from philosophy of mind: you can create a conscious mind in a computer, and what matters is the structure of the computation, not whether the parts are organic neurons. If you had a computation identical to a human brain down to the neurons — a hundred billion of them, wired the same way, with the same synaptic weights, producing the same behavior — Bostrom thinks that would be conscious. Whether you can generate consciousness with less detail is an open question. A loop that only adds one plus one is probably not conscious. A system with a global workspace, attention, and self-representation might be. How much less than a full human brain still counts, he does not claim to know.

Fridman asks whether you must simulate the whole environment or only the brain. Current VR already answers part of it: you do not simulate every detail all the time. Scenes can be procedurally generated and rendered when they enter the player's view. If our environment is simulated, the parts that come into view might be rendered at high resolution while the rest, including every atom in the microphone Bostrom is speaking into, exists only as a coarse representation. Fridman wants the engineering question: how hard is it to construct a reality immersive enough that you would not want to leave, or would not notice you were in a virtual world? Those, Bostrom notes, are two different bars.

Pong and World of Warcraft can absorb people for months without fooling them about reality. The bar for spending an entire life inside, and thereby changing what "reality" means, is higher, and people would differ. Some might refuse a perfect virtual world. Strictly speaking, superintelligence is not required to build ancestor simulations, but if we are in one, Bostrom thinks it was most likely built by a civilization that already had it. The technology to simulate human history at brain resolution is so advanced that machine superintelligence, or biological enhancement of biological brains, would probably have come first.

Robert Nozick's thought experiment asks you to step into a machine, built by future neuroscientists, that can give you any experience you want for the rest of your life: great artist, great lover, a wonderful life, constantly happy. You disconnect from the rest of reality and float in a tank. Nozick thought most people would refuse a permanent checkout, even if they might visit on holiday. He used that refusal against theories on which value is only a function of experience. In the machine you can have any experience, yet many people would say there is not much value in it. So value depends on something besides experience.

Bostrom's sharp observation is how much of that result is a status-quo effect. People may not want to jettison control and plug in. Tell them instead that everything so far has been a dream, and they can disconnect into a real world they know nothing about — or that they are a farmer in Peru growing peanuts — and many might prefer the familiar life of Lex Fridman making podcasts. What we ultimately care about may be projects, relationships, and a self-identity we can carry forward. Tank or soil, Peru or a computer, can matter less than continuity. Fridman says he would take the leap to Peru, especially as VR gets more realistic. Bostrom reminds him that in the experiment, the Peruvian farm is the *real* world, not another virtual one.

Fridman then sharpens the other-people problem. Suppose your wife in the machine has fears and hopes, and when you unplug and plug back in she is still going about her life. Is she real? It depends on implementation, Bostrom says. If she is a static photograph on the wall, tearing her up harms nobody. If she is implemented at neural detail, a fully realized digital mind with your behavioral repertoire, she is very possibly a conscious person, and what you do in the machine has real consequences for how that mind feels. He pushes further: it is not obvious you *can* build an experience machine that delivers a normal human life, including rich interaction with others, without also instantiating those others. A hello can be faked. Days of open, probeable conversation may only be implementable by creating a second consciousness. That, too, he leaves as an open question.

Fridman reports a psychological experiment: he got hold of robotic mobile platforms, made them scream or moan in pain, and found they appeared conscious to him pretty quickly. His introspection says consciousness is easier to fake than intelligence. Features like suffering, hurt, or wondering about existence seem enough to trick a social brain into treating an interaction as real. Bostrom is more cautious. If you do not probe hard, a satisfactory chat may not require you to believe the other party is conscious. If you *do* care, ease of faking becomes the question — and if appearance is that cheap, maybe consciousness is more widespread than we thought, needing only basic intentionality rather than a hundred-billion-neuron brain.

Where appearance and reality come apart, you could create the look of a mind without another mind being there. He stays agnostic about the lines. One observation matters for the simulation argument: how realistic would a virtual model have to be before a simulated creature noticed something was wrong? Consider our own three-pound brains at night. Dreams are immersive, and often you do not realize you are dreaming. If this humble lump of neural matter can throw up a virtual reality that seems pretty real, how much easier would it be for a superintelligent civilization with planetary-sized computers, optimized over eons, to build a realistic environment for you?

Behind that intuition is a claim that brains are not the ceiling of what the universe can build. Fridman floats the opposite: maybe cognitive *constraints* are what enable intelligence, so more powerful computers would be less likely to go superintelligent. Bostrom is unconvinced. Speed alone is a dimension of intelligence, and there is strong evidence you could have computers in this universe much faster than a human brain — perhaps a million times faster, a speed superintelligence. Qualitative superintelligence is harder to argue with a watertight proof, but looking at animals and at Einstein versus a typical person suggests it is not just a matter of being 20 percent quicker. Inventing general relativity is not a slightly longer homework assignment.

Fridman tries to get a ranking. Has Bostrom's thinking evolved toward treating the three cases as equally probable? Bostrom says equal is not necessarily the best assignment, and historically he has punted. Anything that moves one probability around moves the others, because they have to add to one. If almost no civilization at our stage reaches maturity, that is a strong reason to think *we* will fail too. The flip side: as we get closer to maturity, there is less remaining distance in which to go extinct, so fork one looks less likely and the other two pick up probability. The shift need not be monotonic. A new technology that creates a new threat can push the other way — provided the threat can hit civilizations evenly.

Given how little we know, is it even reasonable to put numbers on this? We do not know how hard a simulation is to build, how virtual worlds would change society, or what a theory of everything would do to deep-space exploration. Bostrom thinks a large residual of uncertainty just has to be acknowledged, as with most big-picture questions. We are short-lived, small-brained, cognitively limited, and it is amazing we can figure out as much as we can. Fridman describes the cognitive trick the argument invites: forks one and two *feel* unlikely, so the mind slides to "we are probably simulated." Is that the wrong place to go? Not necessarily, Bostrom says — but the second alternative, about the motivations of technologically mature civilizations, is the one we understand least.

His work on machine superintelligence does not automatically feed fork one either. Losing future value because some process optimizes for a non-human goal would be bad for us. But if we are killed by a machine superintelligence, that machine might still attain technological maturity. The first alternative is not human-exclusive. Any intelligent successor that can reach the ceiling still counts. Replacement could interact with fork two, if the successor were more or less interested in ancestor simulations than we would have been. At first order, death-by-AI is not the uniform filter the first disjunct needs.

The load-bearing fog is motivation. At maturity a civilization may become, in his word, multipotent: able to choose its internal configuration, with enough insight to do so meaningfully. We do not now have direct control over mental states. To feel pleasure you may have to earn money, book a restaurant, and eat. Many instrumental goals would drop out if final goals could be served more directly. Cognitive enhancement, or advice from a superintelligence, would also bring considerations we are now too dumb to see. What looks like a free choice among X, Y, and Z today might then look like an obvious X — or an obvious not-X. He adds that the greatest impact of *individuals* may not be at maturity. It may be earlier, when tracks have not settled, populations are smaller, and indirect effects still fan out.

Suppose fork three is true. Then, counting across the world by the end of time, there would be more simulated people with our kinds of experiences than non-simulated ones. An extra step is required to say *we* are probably simulated. That is an indexical claim about this person right now. Bostrom calls the step the bland principle of indifference: when you have two sets of observers, one much larger, and no internal evidence tells you which set you belong to, assign probability in proportion to the sizes. Ten times as many simulated people with your experiences means you are ten times more likely to be one of them. You do not need infinite time, only however long it takes some civilization, on some planet, to run ancestor simulations — and those starts may be staggered across cosmological epochs.

The Doomsday Argument, first noticed by the physicist Brandon Carter and developed by the philosopher John Leslie, says we have systematically underestimated the chance that humanity goes extinct soon. Most people, Bostrom notes, think something is wrong with it but cannot say exactly what; his own view is inconclusive. The urn analogy is clean. Two identical urns: one with ten numbered balls, one with a million. You pick a ball at random and get number seven. That is strong evidence for the ten-ball urn. Apply the same reasoning to humanity. Consider two hypotheses, 200 billion humans total or 200 trillion. You are roughly human number 100 billion. Under the small total you are a run-of-the-mill person in the middle. Under the large total you are absurdly early. So you should update toward doom soon. The argument needs the self-sampling assumption: reason as if you were a random sample from all humans who will ever have existed, even though you know you are about number 100 billion, alive in 2020.

Rejecting that assumption is tempting, because the implication is so strong from such weak premises. Trouble is, something like self-sampling seems needed for ordinary cosmological inference. In a large enough multiverse, all possible human observations get made, including hallucinations and fluke measurements of the "wrong" constant. We still want to say most observers see the true value, so our measurement of the cosmic microwave background or the fine-structure constant counts as evidence. Without an anthropic assumption, that inference does not go through. The methodological project is to say when observer-selection reasoning is legitimate. Tied back to simulation: the bland principle is much weaker. In the limiting case where everyone with your experiences is simulated, you can deduce you are simulated with no probability at all. As the fraction goes from 90 percent to 99.9 percent, the probability you assign should approach one smoothly, not jump. You can think the Doomsday Argument unsound and still accept the milder step the simulation argument needs.

If physics at the simulators' level resembles the physics we observe, no civilization can perform infinite computation. A positive cosmological constant means the universe accelerates, so even at light speed you can only ever reach a finite sphere of matter. Erasure has a lower bound of loss; protons may decay; you radiate gravitational waves. Eventually there is heat death. Simulations are therefore finite — not zero, not infinite. Stacked simulations are possible, each new spawn with fewer resources, and any observer is uncertain which level they occupy. But every nested computation still has to be paid for by the computer in what Bostrom calls basement reality. Finite basement compute limits how tall the tower can be. If each level adds large overhead, most observers may sit lower down.

Fridman asks about Elon Musk, one of the idea's biggest popularizers. Bostrom's first answer is almost shrugging: if the argument is credible, it tells us something very important about the world, whichever fork is true — arguably one of the most fundamental discoveries on the table. For someone like Musk there is an extra reason. If simulators do not rerun all of history, but only subsets, they may prefer unusual, interesting, or consequential lives. On a scale from a farmer in Peru to Elon Musk, the closer you sit to Musk, the higher that extra boost. What is *outside* the simulation is tangled with what is inside. Views about the creators would help you predict what kind of simulation this is, what happens after, if there is an after. Creatures inside might understand some aspects of their makers and miss the ones that change the meaning of the rest.

He offers a dog-and-human ladder of understanding, then the analogy that explains his famous refusal. Suppose you must guess whether a ten-digit number is odd or even. If even, the best thing in life is to go north; if odd, south. Science and philosophy have given us the first seven digits. We have a lot of information. We are still clueless about the last three, and therefore still clueless about north versus south. That, he feels, is our predicament. We know a lot about the universe, maybe more than halfway to a full picture, but the missing parts could completely change the upshot, including the whole scheme of priorities. Uncertainty is not spread evenly over the surface of the problem. It is concentrated at a few points that decide the map.

After nearly an hour and forty minutes, Fridman turns to superintelligence. Bostrom will not get stuck on definitions. Intelligence is the common-sense bundle: solving complex problems, learning from experience, planning, reasoning. Consciousness is not mixed into it; a system can be fairly intelligent without being conscious. Superintelligence is much more of that — a faster learner, a better reasoner, more effective at achieving goals across a wide range of hard environments. It need not have a human body. Digital is fine. Computers and memories are physical, but the ability to affect the world can be very strong with limited actuators. Typing text on a screen would, he thinks, be ample.

He wants the upside back on the table. *Superintelligence*, published in 2014, focused on what can go wrong because he felt a neglect of what happens if AI succeeds, and a need for a granular map of the pits. Since then, recognition has widened and research groups work on alignment. Public discourse, he suspects, now talks more about negative possibilities than positive ones. He also wants a clean split. Near-term worries — algorithmic discrimination, self-driving cars, drones — are legitimate. The long-term issues in the book are also legitimate. Mix them and you get one bad conversation; keep them apart and you can have two good ones. His diagnosis is overhyping of the near term and underhyping of the long term.

Asked for a positive long-term picture, he finds it almost too obvious to list. The world has many problems, and it is hard to think of one that a friendly, aligned superintelligence would not help with: health, the economic system, investment, trade, foreign policy. There is no single killer app. Artificial general intelligence is the ultimate general-purpose technology. It applies wherever human creativity and problem-solving are useful, which is pretty much everywhere. It would give us much more control over nature. It would not automatically solve problems that arise from conflict between humans. Those are fundamentally political. Some go away with more resources; some still need coordination that more technological capability does not magically produce.

Will there be an intelligence explosion, a phase shift rather than a grind? Bostrom does not know how explosive it will be. It seems fairly likely that at some point AI progress becomes extremely rapid, roughly around human-ish core cognitive faculties — a concept that starts to break down when you look closely. Plotted against other researchers, he would put more probability on an explosion than the average. Once you are in that gray human-level band, is a ceiling waiting? As far as he can judge, a ceiling at or near human capacity is very unlikely. Creating something beyond us is both terrifying and exciting. Fridman hears "lose control." Bostrom does not think that follows. People are working to project higher problem-solving while keeping it aligned with human values. Control is not a given loss.

Fridman restates the sting as specialness: humans would no longer be the smartest thing on Earth. Bostrom's reply is that plenty of things are already smarter, if the universe is big — other civilizations with superintelligences, or brains the size of beach balls. Closer to home, we already live inside systems more capacious than any individual. Treat the scientific community as a mind and it dwarfs any one scientist. Google, Twitter, and Facebook recommender systems are already part of the basic fabric. The search engine is superhuman at retrieving text and images and zero at a vast set of other tasks: a spike of superintelligence, then nothing. Fridman feels AlphaZero is much more intelligent than Deep Blue. Both play board games, but AlphaZero can take a new domain it was not specifically built for, figure it out, and become very good — including, in principle, beating Deep Blue at chess. More general-purpose learning, Bostrom agrees, is what makes a system start to *feel* smart.

What would utopia look like? He does not have all the answers and may never. If the scenario arrives, material constraints expand dramatically and the design space for what we can do, including to ourselves, opens wider than anything in human history. Two things follow. We would need to rethink from first principles what we value — meaning, happiness, fulfillment — not just add a kitchen robot. And we should try to do well by more than one value system. A hedonistic brain in a vat that scores 100 on pleasure and zero on beauty and achievement is the wrong instinct. To a significant though not unlimited extent, you might get 98 percent of the best on several criteria at once. First try to do very well by all of them; only then have the hard conversation about which one gets 97. Happiness and meaning come apart: junk-food reruns can please without meaning, and a costly achievement that helps many people can mean a great deal without being fun. In a low-dimensional space values look tangled; add dimensions and many of them pull apart. Radical abundance is the distinctive context in which some constraints come off. Trial and error, Bostrom has written, is unworkable for existential risk. There is no opportunity to learn from errors. You need foresight, and a willingness to take preventative action and bear its moral and economic costs.